Microsoft trains small LLMs to cooperate in games
Small LMs learn teamwork after failing alone—new training method boosts zero-shot play.
Deep Dive
Key Points
- Small LLMs (1B–7B) trained as multi-agent policies using symbolic 'options' and LoRA adapters, lifting zero-shot performance from 0% to competent play across 3 games
- PA-MAGRPO, a per-agent MARL variant, trains private adapters without gradient sharing, guided by mechanically synthesized feasibility guards
- Behavioral audits reveal reward ≠ cooperation: agents may 'solve' tasks alone while partners idle, necessitating joint behavioral evaluation
Why It Matters
Proves small LLMs can learn true teamwork without manual reward engineering, unlocking scalable multi-agent AI for games, robotics, and beyond.